It is not a theory book. It is a diagnostic for institutions that keep making the same decision, over and over, and calling it progress.

Available now on Kindle and Apple Books. Paperback coming soon.

Issue #006 is live.

Two institutions. Two sectors. The same structural absence.

Knight Capital. 1 August 2012. A software deployment to eight production servers. Seven received the correct update. The eighth was missed. When live traffic hit that single un-updated server, a dormant legacy component reactivated via a repurposed configuration flag. Forty-five minutes. $440 million. The institution could not prove what had been authorised, verified, or running at the exact moment decisions entered the market.

Theranos. 2013 to 2016. A clinical deployment before independent validation. Patients received diagnostic results. Clinicians made treatment decisions. No traceable link between each result and the specific machine that produced it. When CMS arrived, the records that would have made the deployment defensible no longer existed in a form that could be verified.

Both failures. One thing in common. A decision was made and executed. The record that would have made it replayable was never built.

Layer 6 is where judgment becomes a recorded act. Or doesn’t. The distance between explainability and replayability is sharpest here — at the moment of commitment.

Three mechanisms. What they are. Why neither institution had them.

Layer 6 of the Decision Integrity Chain™ — Decision.

Read Issue #006 →

The boundary is gone. Not weakening. Not slowly dissolving while everyone wrote white papers about it.

Gone.

For decades, banks ran as closed systems. Products, risk, servicing, controls - all inside the same corporate wall. Digital transformation meant digitizing the factory. AI meant making the factory faster.

That model is dead.

Today, a single retail banking transaction — one tap, one millisecond — touches technology the bank doesn’t own, data it didn’t collect, models it didn’t build, and infrastructure it doesn’t control. All of it simultaneously. Before any human sees it.

The bank is no longer an institution. It is a node.

Four Shifts Nobody Governed

My earlier research identified four structural forces remaking the institution. (Full analysis: Reimagining the Future of Banking in a Protocol-Driven World). Each has accelerated. None has been adequately governed.

Banks used to compete on proprietary infrastructure - their own rails, their own platforms, their own closed systems. That model is under pressure. Value is beginning to flow across shared infrastructure that nobody owns exclusively. UPI. Onyx. Open finance networks. The banks still defending closed platforms risk fighting for territory that may no longer determine who wins.

Compliance is starting to look different in the institutions paying attention. The direction of travel is clear - rules embedded into the decision itself before it executes, not reviewed after the fact. A handful of banks are moving that way. Most are still producing policy documents that nobody operationalises.

ESG scoring has a credibility problem that greenwashing created faster than any regulator could address. What needs to replace it is harder to fake - ethics built directly into how products are priced and structured. Carbon-aware lending. Dynamic ESG-linked products. Real-time monitoring rather than quarterly reporting. The architecture exists. The adoption doesn’t yet.

Identity is shifting from a compliance file to something more valuable — a trust asset that can be managed and brokered on behalf of the customer. Federated identity, reusable credentials, consent-led models. The banks that move early on this will own a relationship layer that goes well beyond deposits and loans. Most are still treating it as a KYC requirement.

The Unbundling Nobody Governed

Fintech broke the bank apart. Products, channels, data, risk — what used to sit inside one institution now sits across dozens of partners, platforms, and providers.

Then AI arrived and started making decisions across all of it. Fast. Autonomously. At a scale no human oversight structure was designed to handle.

The governance never kept up.

Old-style governance was built for a closed system - clear hierarchy, defined handoffs, someone in the room when the decision was made. That model doesn’t work when the decision is made in milliseconds across systems the bank doesn’t fully own or control.

It doesn’t slow down and adapt.

It just stops working.

Where Does Accountability Live?

Think about what happens when a bank’s automated system makes a bad call. The data came from outside. The model was built by a vendor. The decision executed inside the bank’s own systems. Something went wrong somewhere in that chain.

Who is responsible?

Not the policy document. Not the risk committee that meets monthly. Not the compliance deck that got signed off last quarter.

Most institutions are missing the core distinction. Ownership is structural. Accountability is cultural. But one cannot exist without the other, and in a distributed decision network, most banks have neither.

When decisions are made across multiple systems, vendors, and data sources, nobody automatically owns what happens. Ownership has to be deliberately designed into the architecture. In most institutions, it isn’t. So accountability - which can only form where ownership already exists - never had anywhere to land. That is the design gap. Not a culture problem. Not a leadership problem. A decision architecture problem.

This is what Decision Engineering™ is built to solve. The Decision Integrity Chain™ builds ownership into the system before decisions are made — not after something goes wrong. Because you cannot hold people accountable for outcomes in a system that was never designed to assign responsibility in the first place.

Banks are deploying AI into that unowned space. That gap is where the failures are accumulating.

This stopped being a technology problem some time ago. The question now sitting on every leadership table, whether they’ve named it or not, is simple: Is our institution designed to own its decisions, or are we just hoping accountability works itself out?

The danger was never the fintechs. It was always irrelevance. And the reward was never market share. It is trust-share.

The institutions that see this now will architect for it. The ones that don’t will regulate their way through the wreckage.

When regulatory compliance buckles inside a distributed decision network — does it fail at the protocol layer, or does it fail because governance was never engineered in to begin with?

© Deepak Aggarwal, 2026. All rights reserved. Decision Engineering™ is a trademark of Deepak Aggarwal. Reproduction or use without written permission is prohibited.

This post was inspired by ’s recent piece on deposit displacement, which nails the visible part of the story. The $3.15 trillion that left banks for fintech accounts between 2020 and 2025 is real. The paycheck motel argument holds.

But there’s a second problem sitting underneath that one, and nothing in standard reporting picks it up.

Deposits don’t have to leave to become unstable.

The bank that looked fine

Earlier this year, I published on SSRN a composite regional bank in the Asia Pacific, patterns drawn from real observed data. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6147568

On the metrics, this bank was fine. LCR comfortable. NSFR above minimum. Deposit balances holding flat. No run, no regulatory concerns, credit book clean.

And yet the SME and corporate operating accounts were changing. Not because clients were under pressure - they weren’t. Rates had moved, and money market funds were offering 100–150 basis points more than the bank’s operating accounts. The treasury software was doing the rest: idle balances swept out each night above a set threshold, back in by morning to cover the day’s payments. No call to the relationship manager. Just a rule, running on a schedule nobody at the bank had written.

Screen balance: flat. The money itself is gone by 6 pm, back by 8 am.

The analogy that fits - a tenant paying rent every month on time, but who quietly sublet the flat six months ago and only comes by to collect the post. Dashboard says occupied. The flat’s been empty most nights.

What that flat balance was hiding

The loans were still on the books. Still needed funding. But the deposit money the bank assumed was backing them, not reliably there anymore after the close of business. So, wholesale markets filled the gap. Interbank borrowing, institutional money, all of it priced higher. Wholesale funding costs moved from roughly 3.6% to above 4.4% inside two quarters.

Loan book: unchanged. Liability mix: quietly not what it was.

LCR stayed green. NSFR stayed green. No alarm.

Funding quality had deteriorated - not from deposits leaving but from the deposits that stayed, stopping from behaving like stable funding. That distinction matters. Most governance frameworks weren’t designed to see it.

Why nobody caught it

Not a failure of any single team. A chain failure.

The strategy had been built for a deposit base where money moved when a client decided to move it - a deliberate act, visible, slow enough to manage. That deposit base no longer existed. Treasury platforms with API connectivity and automated sweep logic had changed how corporate and SME liquidity was managed, and the strategy had never been updated to reflect any of it.

Rules classified SME operating accounts as stable funding. Product label: operational account. Nobody had formally checked whether the behavior still matched that label. It didn’t.

Across the functions, the individual reads were accurate. Treasury saw a pricing competition problem. Business lines saw active, transacting clients. Risk saw clean credit. What none of them owned and what the architecture didn’t require anyone to own - was the question underneath all three: Is the funding base actually behaving the way the models say it is?

It wasn’t. But asking that question wasn’t anyone’s job.

The feedback loop ran on LCR and NSFR. Both compliant. Behavioral signals - sweep velocity, rate sensitivity, effective duration compression in the SME book - never entered the formal governance cycle. Drift had been accumulating for quarters before a supervisor named it.

Every layer functioned as designed. That was precisely the problem.

The Fiduciary Gap

There’s a name for this kind of structural divergence: the Fiduciary Gap.

The institution was optimized toward regulatory ratio compliance. Its actual obligation was something different - sustainable funding, money that would genuinely be there when the loans it backed needed servicing. For years, those two things aligned. Then, without anyone deciding it, they stopped.

No rule broken. No individual decision wrong. Ratios green while the behavioral reality underneath shifted quarter by quarter.

It surfaced in a routine supervisory review. One question from across the table:

“If these balances are operational, why do they change daily in line with money market spreads?”

Quiet room. No violation to point at. No clean answer either.

That’s what the Fiduciary Gap looks like when it surfaces - an institution that had been reporting stability it hadn’t verified, not dishonestly, but because the assumption had never been formally re-examined. The old definition of stable funding had stopped being accurate some time ago. The governance architecture hadn’t noticed.

If you’re inside a bank reading this

One question worth sitting with: do your deposit stability numbers reflect how clients actually manage liquidity in 2026, or how they managed it before treasury platforms could connect to bank accounts via API and sweep balances on a nightly rule?

Not the LCR. The behavior behind it.

Are your SME and corporate clients running automated sweeps you don’t see? Can your treasury team tell the difference between a balance that’s genuinely sticky and one that’s clearing payments in the morning and sitting in a money market fund by evening? Is that distinction in your stability modelling at all?

If those answers aren’t in the data, that is the finding.

The retail displacement story Ron documented is visible. Bankers know it. The behavioral reclassification of SME and corporate operating balances is the one that doesn’t surface in outflow reports - it surfaces in funding costs, in NIM compression, in wholesale dependency, typically a year or more after the behavior causing it started running.

Boards that identify this early get to make a deliberate strategic call - which balances to defend, at what pricing cost, with what funding trade-offs. The ones that find out from a regulator in a quarterly review don’t get that conversation first.

Stability isn’t a classification decision made at account opening. It’s something that has to be re-earned continuously, against actual observed behavior - not assumptions that made sense a decade ago.

This post draws on my January 2026 SSRN paper “When Stable Deposits Stop Being Stable: A Behavioral Reassessment of Core Funding in Regional Banking.” The Fiduciary Gap and Decision Integrity Chain™ are proprietary components of the Decision Engineering™ framework.

© Deepak Aggarwal, 2025–2026. All rights reserved.

Issue #005 is live.

Two institutions. Same country. Same period. The same structural absence.

Apple Card. Gender wasn’t an input. Variables that correlated with gender were. Women with equivalent credit scores and assets got different outcomes. Goldman Sachs had never defined what the model was not permitted to do.

UnitedHealth deployed nH Predict to determine post-acute care coverage. The algorithm predicted a length of stay. Reality exceeded the prediction. Coverage was denied. Didn’t matter what the physician documented. Nine of ten appealed denials were reversed. February 2025 - a federal judge allowed the case to proceed to trial. March 2026 - the court ordered disclosure of the algorithm itself.

Both failures. One thing in common. A model with a function and no boundary. Layer 5 is where that boundary should have been defined. It wasn’t.

Three mechanisms. What they are. Why neither institution had them.

Layer 5 of the Decision Integrity Chain™ - Judgment.

Click here to read Issue #005

Issue #004 is live.

Two institutions. Two sectors. The same structural failure.

Wirecard had a regulatory classification. It also had a balance sheet that had long since outgrown that classification. For years, nobody asked whether the rules still governed the institution they were written for.

Orpea had a care mandate. It also had 1,100 facilities across 23 countries. The rules set minimums. The minimums were met. The institution those minimums were designed for had ceased to exist.

This issue examines what happens when rules are followed precisely - and the outcome is catastrophic anyway. And what three mechanisms would have changed.

Layer 4 of the Decision Integrity Chain™ - Rules.

Read Issue #004 →

Issue #003 is live.

Two institutions. Two sectors. The same structural failure.

Wells Fargo had a Purpose. It also had a cross-selling mandate. For six years, nobody defined what that mandate would not permit.

Kaiser Permanente had a clinical mission. It also had an integration mandate. The mandate delivered exactly what it was designed to deliver. The floor it needed to protect patients was never set.

This issue examines what fills the space between a mandate and its boundary, and what three mechanisms would have changed.

Layer 3 of the Decision Integrity Chain™ - Intent.

Read Issue #003 →

Issue #002 is live.

Two institutions. Two sectors. The same structural failure.

Credit Suisse had a Purpose. It also had a growth strategy. For eight years, nobody formally asked whether the two were still compatible.

NHS England had a clinical mandate. It also had an Elective Recovery strategy. The strategy delivered exactly what it was designed to deliver. The Purpose it was supposed to serve was never formally placed above it.

This issue examines what that costs — and what three mechanisms would have changed.

Layer 2 of the Decision Integrity Chain™ - Strategy.

Read Issue #002 →

Issue #001 is live.

Two institutions. Two sectors. The same architectural failure - and the three mechanisms that would have prevented it.

This issue introduces the Decision Integrity Chain™ and its first layer - Purpose. Every issue that follows moves one layer deeper.

Read Issue #001 →

Modern banking doesn’t fail because people don’t do their jobs. It fails because no one owns the decision end-to-end.

Design, validation, and distribution operate in silos. Each function is optimized. Each control is satisfied. Each metric is met.

But the decision itself is never held together as a single chain. So, it gets assembled.

And when a decision is assembled instead of owned, the risk doesn’t disappear.

It moves.

To the client. And eventually, back to the institution’s reputation.

The Case Study: The 2021–2022 Autocallable Collapse

In late 2021, banks issued billions in autocallable structured notes linked to “high-growth” stocks like Lucid, DocuSign, and Chewy.

The Pitch: Earn ~10% annually, as long as the stock stays above a “knock-in” barrier set at 50% of the initial price. A 50% buffer. It sounds like a fortress. Simple. Intuitive. Reassuring.

The Illusion: That buffer was calculated at the absolute peak of the market. For a stock like Lucid, trading above $50, the barrier was set near $25.

The Reality: By early 2022, the stock had fallen below $20. The “buffer” had effectively disappeared before the client even understood what it was protecting them from.

The Anatomy of a Flawless Disaster

One autocallable issued in January 2022 returned less than $10 for every $100 invested. A 90% loss.

This didn’t happen because the product failed. It happened because it worked exactly as designed. Every function did its job.

Each layer was internally consistent. But the assumptions did not travel across the layers. The “50% buffer” was mathematically correct but economically meaningless. It was 50% of a bubble, not 50% of a realistic trading range.

The model was right. The interpretation was fragmented. The outcome was inevitable.

The Stability Fallacy

This is not just a structured product problem; it is an architectural one.

In 2023, when banks came under stress, the fastest-moving capital didn’t come from retail depositors. It came from Wealth Management clients - the very segment every private bank models as their most “sticky” and stable base.

The theory said, “sticky.” The client said “goodbye.”

The discrepancy wasn’t behavioral. It was architectural.

It was a failure to recognize that the decision architecture - the way the bank understands and communicates with its most sophisticated clients- was broken.

The Strategic Pivot: From “Correct” to “Owned”

Regulators are already shifting their gaze. They are moving away from: “Are your models accurate?” and toward: “Who owns the decision your models are making?”

Correctness at the component level is no longer enough. Banks have mastered how to design products and validate models at scale. What they have not built is the layer that ensures the decision makes sense as a whole.

The Real Question: Who sits between the Quant and the Relationship Manager? Who, inside the institution, is accountable for the full end-to-end decision?

In most banks, that role does not exist. And when it doesn’t, the institution executes flawlessly, the controls all pass, and the client absorbs a loss that no one inside the system would have approved if they had seen it end-to-end.

The Head of Business carries accountability. But accountability for outcomes is not the same as ownership of the decision as it moves through the institution. By the time it reaches them, it has already been assembled.

In a world of automated execution, the only true competitive advantage left is the integrity of the human decision that started the chain.

Note on Sources & Use All examples draw on publicly available disclosures and market data to highlight recurring structural patterns. This is not investment, legal, or regulatory advice. © 2026 · All rights reserved.

Insurance does not break when a bad decision is made. It breaks much earlier.

At the point where no one can see the full decision anymore.

Not because the data is missing.

Not because the models are wrong.

But because everything sits in different places, and no one owns the whole picture.

This is not a theory. Ask Lloyd’s in 1992. Ask AIG in 2008. Ask every California carrier in 2018.

For a long time, this doesn’t show up.

Numbers look fine. Capital looks strong. Losses are absorbed.

So, it feels like things are working.

Until one simple question becomes difficult to answer:

Do we actually understand the risks we are taking, or are we just looking at pieces of it and assuming it adds up?

The Illusion of Health

The three years between 2023 and 2026 have reshaped global insurance. The industry survived the soft landing. What it landed into is far more complex than the one it left.

If you are sitting inside a carrier right now, the numbers tell a reassuring story.

• Return on equity has recovered to ~10%

• Investment income is now a primary earnings driver after a decade of suppression

• Reinsurance capital stands at a record USD 838 billion

• Balance sheets look clean. Capital looks abundant.

• Six consecutive years of >$100 billion insured losses have been absorbed.

Based on traditional metrics, there doesn’t seem to be a problem.

But behind those numbers, tens of thousands of decisions are being made each day.

Whom to cover. At what price? On what information? Based on what assumptions? Approved by whom? Countermanded by whom? And for what reason?

None of this is model output; this is judgment calls. And judgments can be fragmented, even where the result is clear.

In California, insurance companies modeled risk based on three decades of data. All the models agreed, all the assumptions were correct, all of which came from a world no longer existent when payouts began.

This was not the problem with the model; it was the problem of using the model without challenging its world. No one owned the assumption that the past would reflect the future. And by then, it was too late, because the policies had been written, the fees paid, the capital allocated.

But numbers reflect results. Numbers don’t reflect decisions. Understanding what’s going on at an institution means digging down one layer.

Let’s take one.

The Decision That Doesn’t Exist

Now take a single, ordinary policy decision.

A commercial risk that is quoted digitally, reviewed through underwriting, and adjusted via distribution.

Same customer. Same exposure. Nothing new or different.

Different views at each step.

• Pricing sees a risk signal but not the customer.

• Underwriting sees the customer and not the pricing logic.

• Distribution sees the relationship, but neither the customer nor the pricing logic.

It was approved. No rule has been violated. No control has been bypassed.

But the only question worth asking is: What was the real risk undertaken by the organization?

Nobody had all the facts together. Nobody made all the judgments together. It was just the coming together of a series of partial judgments leading to the approval.

This is not a decision. It’s an outcome with no ownership.

This happens everywhere. In every institution. On every ordinary day.

That is not a comfort. That is the problem.

The Structural Debt No One Prices

This is where the industry is slowly and silently accumulating risk, while being healthy on the surface.

The risk is not being shown on the balance sheet through data. But it is inside the decision process itself.

The Governance Gap

It’s well known that AI is fast moving up the chain from pilots to core operations. In 2025 alone, AI-related deal activity in insurance surged 328% in value. Execution has scaled. But when MAS, RBI, and ECB ask carriers to explain a specific decision end-to-end, not the model, the decision, most firms do not have a clean answer. That gap is not a technology problem. It is an accountability problem.

The Protection Gap

Within ASEAN, insurance penetration still sits at ~3.2% of GDP compared to ~7% globally.

This difference has endured despite capital, innovation in product design, & innovations in distribution. And it continues due to the inability of institutions to build a full picture of the risk they are asked to underwrite.

Consider where real growth is coming from - parametric products. Products that do not underwrite the entire risk but rather rely on observable parameters. Rainfall, not crops. Wind speed, not buildings.

This is not innovation. It is simply a workaround.

A workaround on a large scale means the root cause of the problem has yet to be solved. The protection gap is not a distribution issue. It never was. It is a decision construction challenge.

The Explainability Illusion

Regulatory direction is converging globally on decisions and not on models.

The question is no longer: “Is your model accurate?”

It is: “Can you explain what actually happened?”

The Paradox

Most enterprises see bad decisions impact them immediately. In insurance, the decision fragments earlier than anyone notices, and the consequence arrives later than anyone expects. The gap between the two can be years. Sometimes decades. And in between - the numbers look fine.

A life product. A long-tail liability. Climate risks underwritten today but materializing in 2031.

By the time you realize your decision was wrong, the capital allocation has happened, the premiums have already been paid, and reserves have been set based on assumptions that were never tested.

Insurance capital is more abundant than at any point over many years. Yet, more uncertain than it acknowledges with regard to the risks they underwrite.

$838 billion in reinsurance capital is built upon:

• Assumptions about diversification

• Assumptions about correlation

• Assumptions about exposures

But assumptions are made in the context of coherent decisions.

When decisions become fragmentary, the assumptions themselves are weak.

Capital, instead of providing a layer of protection, becomes a statement of confidence based on unfounded reasoning.

The bigger the capital pool, the bigger the assumption underpinning it.

The bigger the assumption, the greater the risk exposure in case of failure.

That’s the paradox.

The Only Question that Matters

“Which model should we use?” – the industry continues to ask.

But that is already an outdated question to ask.

The actual question is much simpler and more difficult:

Can you map out one entire decision from end to end, covering all models, all data sets, and overrides, and explaining why it was a good decision to make at that time?

And not in theory, but in practice, under scrutiny, on demand.

If the answer is negative, speed won’t be your edge.

It will be how fast you can grow your indefensible decisions.

I have been mapping out precisely where the process fails. Not at the model level, but right when a decision becomes incomplete.

This is the new frontier of risk.

If you found the thread in your institution, I would like to hear where it broke.

The institutional examples referenced in this post - Lloyd’s, AIG, and the California wildfire market, are cited as publicly documented historical events for illustrative purposes only. This post does not constitute legal, regulatory, or investment advice. All views are my own.

© Deepak Aggarwal, 2026 · The Decision Architect · Original IP · All rights reserved

It is Thursday morning in Nashville.

Limestone Financial, a $28 billion regional bank with a strong SME business, a growing payments modernization program, and a board that takes governance seriously, is holding its quarterly Q3 2025 audit committee meeting.

There are three reports on the table.

Sarah Chen, Chief Treasury Officer, is first up.

Deposits are holding up well. LCR is at 136%. Funding costs are up but within guidance.

“We are seeing some pricing pressure in the SME segment,” she says.

“But balances are stable.”

Next up is Marcus Webb, Head of Payments and Digital.

The instant rails are live on 80% of the commercial book. Transaction volumes are up 23% year on year. AI-assisted fraud decisioning is reducing false positives. And processing time is down 60%.

David Okafor, Chief Risk Officer, is last up.

“No credit deterioration,” he says. “Models are performing within tolerance.” “And the new AI-powered module for underwriting is running ahead of plan.”

The committee chair looks around the table.

Everyone is green.

“Good work, everyone,” he says.

And with that, the meeting is adjourned.

Forty-seven minutes.

Nobody asked the question that connects all three reports.

What SVB Actually Taught Us and What We Haven’t Applied

Two months later, the CFO at Limestone is on the phone with an analyst, who asks a question about the Q3 2025 data, which is running lower than the firm’s expectations.

“What’s driving the rise in funding costs? And why is the net interest margin compressing so hard?”

The CFO knows the numbers. Liquidity ratios are high. Deposit balances have stabilized. Everything is well within the bounds of the policies.

What the CFO cannot immediately answer is why the marginal cost of funding is rising so abruptly, even as the aggregates appear stable.

The balance sheet is in good shape. What’s changed is the behavior behind it.

In the time between reporting cycles, the operating balances of SMEs have become more rate-sensitive. Excess funds have become more quickly, more automatically, and more frequently deployed than the funding model anticipated.

Deposits have not fallen out; they have fallen in.

By the time the aggregate data presents a board-ready story, the impact of the behavioral duration compression is already manifest in the wholesale funding profile and the margin pressure.

Nothing is wrong. The assumptions are simply lagging behind the behavior.

When SVB collapsed in March 2023, $42 billion was withdrawn in hours. Not because anything was wrong with the fundamentals. Not because anything was wrong with the institution itself. Because no one had bothered to build a digital infrastructure that would slow the panic down a little.

Because customers did not stand in a queue outside the branch. They stood in a queue on their phone. And the information that caused the queue to form was shared in a group chat before the board had a chance to read the briefing note.

Because the risk management system was sound. Because the assumption about the speed at which customers would act was not.

Most regional banks in the US looked at SVB and realized that they need to have better liquidity buffers and better monitoring of uninsured deposits.

LCR was raised. NSFR was raised. Uninsured deposit ratios moved from secondary disclosure to boardroom and earnings call topics.

These were the right actions to take from an operational standpoint. However, the question of whether the behavioral assumptions underlying the funding model were valid in a world where money moved instantly was not given the same degree of attention.

The distinction matters because operational fixes address what happened. Behavioral assumption updates address what will happen next time, under different conditions, at a higher speed.

What Changed in the Deposit Base

Back in Nashville. Sarah Chen is having an even deeper discussion with her team. The aggregate figures look good. But Sarah has also pulled the behavioral duration analysis by segment for the first time in two years. What she’s seeing isn’t comfortable.

The SME operating balances that were modelled on stable funding over three years are behaving as if they are six-month money. They haven’t moved; the commitment behind them has.

I have written about this before in the context of Asia Pacific banks. The behavior is now evident in the US market.

The Federal Reserve Bank of New York published research in July 2025 that revealed deposit flightiness had reached historic highs as a result of quantitative easing and that this high level of flightiness would magnify the banking sector’s response to subsequent rate increases. The online savings rate had risen above 4.5% in early 2025 compared to the average of 0.47% for brick-and-mortar banks, a gap that automated treasury systems are specifically designed to take advantage of. And the regulators have specifically highlighted that corporate depositors with higher balances are significantly more rate sensitive than other depositors, and that assumptions over the decay rate have yet to be updated to reflect post-2023 behavior.

The data is there. The question is, have the assumptions in your funding model caught up with the data?

SME and corporate operating balances are no longer passively sitting in an account, waiting for someone to move them. They are being actively managed, in real time, automatically, versus yield alternatives that are immediately visible and immediately accessible.

Treasury management platforms, API connections to bank accounts, automated sweep triggers, and formalized cash diversification policies, many of those policies written after 2023, specifically limit exposure to individual banking partners.

The relationships are still intact, the accounts are still open, and the transactions are still flowing.

But the surplus is no longer sitting in the same place, and it will continue to move, rationally, efficiently, automatically, as long as the spread is attractive.

These are illustrative benchmarks derived from publicly available supervisory research, FDIC call report trends, Federal Reserve Bank of New York working papers, and industry observations. The figures provided represent a range for the US regional banking cohort as a whole and should not be attributed to any specific institution. It is recommended that you review against your specific institution’s data as well as any applicable supervisory guidelines. This table should serve as a tool to generate questions for internal diagnostic purposes only.

This is not crisis management; this is optimized treasury management, and this is now structural.

At Limestone, everything looked okay, nothing out of the ordinary, until the team dug in and found that four things had changed in the two quarters leading up to the crisis:

• Behavioral duration in the SME segment had compressed by nearly 40%,

• Wholesale funding had stepped in to cover the gap.

• Marginal funding costs had risen by nearly 80 basis points.

• And every regulatory classification had stayed the same.

The metrics said safe. The assumptions didn’t.

What Payments Modernization Added

Marcus Webb is proud of what he and his team have built. And he should be. The instant rails are working exactly as intended.

But at 2 am on a Tuesday, an algorithm detects a strange occurrence in a group of SME transactions and reacts by reducing credit limits for 340 accounts. By 6 am, seventeen small business owners are unable to pay their staff.

By 9 am, Marcus knows something is amiss. By 11 am, Risk knows too. The board is informed at the end of the week.

The algorithm wasn’t wrong, just quick. And the governance, well, that was operating at Thursday morning meeting speed.

We’ve spent a decade making money go faster. Tap to pay, instant transfer, real-time settlement.

It’s worked as designed.

What we didn’t redesign was how decisions move.

A fraud rule that was too restrictive was used to incorrectly stop a handful of transactions over the course of a night. Someone would notice. We would fix it in a day or two.

The same flawed rule today can incorrectly stop thousands of payments in minutes. Merchants can’t settle. Customers can’t get what they want. Social media goes crazy. And the institution’s response still operates at the speed of our Tuesday morning meeting.

A credit algorithm that adjusted limits too severely was used to incorrectly affect a set of transactions processed over a night. Now it can incorrectly adjust limits on thousands of small businesses in real-time because sales were down for a few hours. A weather event has become a cash crisis before anyone ever looks at the code.

The rails have been rebuilt for speed. The brakes have not.

And the speed of money versus the speed of our institution’s response is where institutional risk resides today, silently, invisibly, and growing exponentially.

What AI Is Accelerating

David Okafor’s risk dashboard is still green. The AI underwriting module is operating within tolerance. Default rates are within model parameters.

However, after six months, no one has asked whether the decisions the model is making today are still consistent with the intent of the governance that authorized the model twelve months ago.

The model has been optimized. Quietly. Efficiently. Legally. One threshold at a time.

The policy document has not changed. The implementation has.

Here is how this process usually works: A credit threshold is shifted by half a point because the model’s performance metrics were slightly improved. Then another threshold is shifted, and another weighting factor, and so on. Each of these changes was reasonable in isolation, and each of these changes was within the model’s authorized operating parameters. However, after six months of continuous optimization, the population of borrowers that the model is now willing to lend to looks very different from the population of borrowers that the credit committee originally had in mind when they signed off on the model’s original deployment. No single person made a bad decision; the aggregate drift simply was not reviewed.

AI-assisted decision-making is exponentially increasing the governance gap.

This is not an argument against AI in banking. The efficiency improvements are real. The cost savings are measurable. The processing enhancements are genuine.

However, every time an AI system replaces a decision that a human made, two things must be true. The logic of the AI system must be consistent with the governance intent of the institution. And someone with authority must be able to answer, in real time, why a particular decision was made.

In most institutions today, neither of these things is necessarily true.

The AI is functioning. The metrics are within acceptable limits. The dashboards are green.

But the question that a supervisor will eventually ask, indeed, that the OCC and Federal Reserve are already asking in the US, is not whether the model is functioning. It’s whether the decisions the model is making are still aligned with the intent of the original authorization.

When a model continuously optimizes, when human review effectively becomes pro forma, and when the original policy document remains unchanged, well, the institution has changed. Quietly. Not anybody’s fault. One small optimization at a time.

The problem won’t show up in the metrics. It will show up under stress.

The Question Nobody Asked

It’s now six months after that Thursday morning in Nashville. Limestone is in a supervisory review. The examiner poses a question that freezes everyone in the room:

“Walk me through how a specific SME credit decision made last Tuesday was authorized – from the governance intent at the board level to the model output to the customer outcome.”

Sarah, Marcus, and David glance at each other. They each know a part of the answer. They each know a part, but none know the whole answer.

That’s not a compliance problem. That’s an architecture problem.

And here are all three connecting threads.

1. Deposit behavior is changing. It’s becoming faster and more rate-sensitive.

2. Payment systems eliminated response time.

3. Artificial intelligence sped up the decision-making process beyond what governance was intended to review.

Each of those things, individually, is something that can be managed. Collectively, they produce a compounding effect between how fast the environment is moving versus how fast the institution is governing.

The audit committee that met on that Thursday morning in Nashville was not negligent. Sarah was doing her job correctly. Marcus was doing his job correctly. David was doing his job correctly. The board chair was doing his job correctly.

What was not being done was: are these assumptions underlying these three clean audits still valid in a world where all three of them intersect?

That is the question of governance. And it is a question for the board.

Not because the board needs to understand AI models, not because they need to understand payments architecture, not because they need to understand deposit beta analysis.

It is because the only place within the institution where all three of those conversations intersect is at the board level, and where someone has the authority to say: are we governing the speed of change within our environment, or are we simply reporting on it?

The Architecture Fix

To manage the pace of change, there must be an internal architecture that can connect these silos in real time. To close this gap, there are three upgrades to the institution’s “wiring” that are required to close this gap:

• Closed-Loop Traceability: A system that can identify when there is a drift in the AI model that moves the profile away from the original Board-approved governance intent.

• Algorithmic Circuit Breakers: Triggers that can “pause and notify” humans in real time when payment or credit decisions hit a certain level of volatility.

• Cross-Silo Indicators: A single “Speed of Change” indicator that can monitor the intersection of deposit flightiness, payment velocity, and decision drift

Three Questions for Your Next Audit Committee Meeting

1. If your largest SME customers and corporate depositors are using automated treasury systems, what does your behavioral duration analysis actually show for them by segment, and when was the last time you updated the stability assumptions within your funding model?

2. If your payment infrastructure now processes funds more quickly than your risk review process, what is your response mechanism when an algorithmic decision is wrong? Who is ultimately responsible for that decision, and how quickly can you unwind it?

3. If your AI-driven decision processes have significantly reduced human review, what evidence do you have to support that the decisions now being made are still aligned with the intent behind the governing decisions to implement these systems, and not simply with the metrics?

If you can’t answer these questions well, it’s not a failure. It’s a beginning.

Limestone Financial is an illustrative institution only, but the things that are being talked about are real.

This article is part of a series I’m writing about decision-making in financial institutions, the gap between what people want and what actually happens in financial institution governance, and what to do about it. Some of the previous articles in this series are “When deposits stopped behaving the way models expected” (December 2025) and “Your Modern Payments are Moving at Light Speed.” Your Risk Strategy is Still on a Tuesday Morning? (February 2026).

The behavioral deposit analysis is derived from my SSRN Working Paper titled “When Stable Deposits Stop Being Stable: A Behavioral Reassessment of Core Funding in Regional Banking” (January 2026).

The governance issue isn’t whether your institution has the right frameworks in place. It’s whether the distance between what the governance model was designed to do and what your institution is actually doing has grown so large that it matters, and whether anyone is currently capable of measuring that difference. This is a harder process than simply looking at a dashboard, but it’s a lot more valuable.

I help boards and senior management teams do exactly that: measure where the assumptions around deposit behavior, payments responses, and AI decision drifts have taken the institution beyond what the governance model was ever intended to address. No frameworks sold. Just a clear understanding of where the drift is occurring, what it means, and what it means to your institution.

We have spent billions making money move at the speed of light. However, we forgot that humans still think at the speed of a Tuesday morning meeting.

For years, payments modernization has been sold as a pure tech success story:

• Tap and pay.

• Instant transfers.

• Refunds in minutes.

• Loans approved at checkout.

It feels efficient. It feels smart. And in many ways, it is. But a dangerous gap has opened up: Money now moves faster than businesses can think.

When the “rails” move at light speed, but your “brakes” (risk and governance) move at human speed, the results aren’t just glitches, they’re systemic crises.

What has actually changed?

Real-world evidence: Why speed is a double-edged sword

1. The Instant Bank Run (Silicon Valley Bank)

There were no lines outside the building when SVB failed. Customers weren’t lined up on the sidewalk; they were lined up on their phones.

• Reality Check: $42 billion fled the bank in a matter of hours.

• Lesson Learned: The bank’s technology was cutting-edge, but its risk management strategy was not designed to handle money fleeing that quickly. When money flows instantly, panic flows instantly too.

2. The “Rainy Day” Algorithm (UPI in India)

Banks can now monitor every transaction in real-time. This is a very impressive-sounding feature, but it gives rise to a “sensitivity” issue for small businesses.

• The Situation: A small business has two slow days due to a rainy day. Sales are down. An algorithm interprets this as “increased risk” and reduces credit limits instantly.

• Lesson Learned: The business owner didn’t become riskier; the weather changed. A slower system would have “smoothed out” this issue. Now, every small change results in an immediate, knee-jerk reaction.

3. The Instant Refund Culture (Visa Direct / Mastercard Send)

Now, platforms can instantly refund money to customers. Good for customers, but a shock to the seller.

• The Scenario: The seller watches money leave their account in minutes based on an automatic claim.

• The Lesson: “Review time” used to exist. Now, the process is judged first, and the reasons come later. As speed rises, review time vanishes.

4. The Death of the “Float” (PayNow in Singapore)

Sellers adore instant money, but they have no “cushion” anymore.

• The Scenario: The float was the time between payment and receipt. It was the shock absorber for liquidity.

• The Lesson: Efficiency rises in good times. In bad times, the pressure rises 10 times faster since there is no “transit time” to hide behind.

So, what’s the Real issue?

One may think the issue is that fast payments are bad. No, it’s not.

While businesses modernized the payment platforms, they didn’t redesign how they make decisions.

For example, if a fraud rule were too strict, a few transactions would get blocked. Someone would notice. The team would fix it in a day or two.

But now, the same wrong rule can block thousands of payments in minutes. Customers get locked out. Merchants can’t settle. Social media lights up.

Nothing malicious happened. Just speed.

Take another example: If a credit algorithm tightens limits too aggressively, it might affect a batch processed overnight. But today, it can shrink limits across thousands of small businesses instantly because sales dipped for a few hours. A short-term blip becomes a cash crisis.

That’s the real shift.

We have modernized how money moves, but we have not yet modernized how decisions move.

“Justify”: To whom, what, and why?

Someone must be able to respond if a system automatically withdraws money, tightens a credit line, or freezes an account:

• What specifically set it off?

• Based on what proof?

• In what authorized policy?

• Did the reaction match the level of risk?

What if this occurs on a large scale? Who takes ownership of it?

• Operations?

• Risk?

• Technology?

• The Board?

Speed is impressive.

But institutions are not trusted because they are fast. They are trusted because they can explain themselves.

The Real-Time trap

The rails have been modernized, but the brakes have been overlooked.

These days, a bad algorithm doesn’t just result in a “glitch”; it quickly destroys hundreds of small businesses. Before the Board can even make a Zoom call, a liquidity rumor causes a complete bank run, not a line at the door.

In summary, real-time accountability is necessary for real-time money. You are inefficient if your systems respond in milliseconds, but it takes you three days to figure out “why.” You are vulnerable.

The speed is amazing. However, explanation, not just speed, is the foundation of trust.

Most leaders think that Banking, Healthcare, and Logistics are different domains. They aren’t.

Regardless of whether you are running a $500M credit book, a surgical department’s patient throughput, or a worldwide shipping route, you are trying to solve the same mathematical problem: Optimizing outcomes under high-stakes uncertainty.

In my years of experience in Banking and Healthcare, I have come to understand that the “industry” is merely the packaging. The core is Decision Quality (DQ).

Decision Quality is not a marketing tagline. It is an engineering practice.

I call it Decision Engineering: designing how information turns into decisions, and how those decisions are owned.

Across industries, failures don’t happen because there’s no data. They happen because information doesn’t turn into action properly.

Every system is built to move something.

• In Banking, it moves money.

• In Healthcare, it moves patients through the care process.

• In Logistics, it moves goods from the origin to the destination.

When that movement slows or stops, pressure builds. Risk compounds.

Different industries. Same structural problem.

Things stop moving when decisions are slow, disconnected, or wrong.

Today, we are headed towards Agentic Decision Systems. We don’t merely require “data-driven” decision support systems. We require decision-focused system design.

Dashboards inform. Agents act. But acting without architecture multiplies fragility.

This is where the Decision Integrity Chain™ is so important:

Purpose → Strategy → Intent → Rules → Judgment → Decision → Outcome → Feedback

Whether you are underwriting credit, allocating ICU beds, or redirecting shipping containers, this chain rules the transition from information to action.

As geopolitical tension alters trade routes, interest rates change, or supply shortages cascade through Tier-N suppliers, it is not the institution with the most information that will survive.

It is the one that has engineered its Decision Integrity Chain to be robust under stress.

The same stochastic models that forecast credit defaults can be repurposed to rate the risk of Tier-N suppliers. The math is transportable. The architecture is what matters.

What most contemporary supply chains still fail to grasp is this:

• They have invested in visibility, but not in decision-making accountability.

• They can see disruption in real-time.

• They can track containers down to the mile.

• They can map the supplier risk across tiers.

But when rerouting decisions need to be made, when capital needs to be redeployed, when contract risk needs to be absorbed, ownership gets fuzzy.

Analytics are crystal clear. Decision-making authority is not.

In many companies, data travels faster than decisions. And that’s where vulnerability lies.

The lesson from banking and healthcare is simple:

Systems fail not because things stop being. They fail because things stop flowing.

• Dependencies need to be understood.

• Escalation routes need to be clear.

• Decision ownership needs to be clear.

• And the way from information to action needs to be designed.

These rules do not change when dealing with supply chains. The variables are the only ones that change.

These remain true for supply chains as well. Only the variables change.

Resilience is not redundancy. It is Purpose-to-Feedback alignment.

I am currently writing out my Decision Architecture framework on this Substack platform as I work with global companies to help them stabilize complex, high-risk systems.

The question today is not whether you are “AI-enabled.”

It is this:

Are you building a Data team?

Or are you building a Decision Engine?

Because in high-risk systems, the answer to this question is what will make the difference between the flow surviving stress or failing under it.

Executive Brief

The Problem: Private Banks are growing at an unprecedented rate, yet growing “Forgetful.” We are witnessing a Context Vacuum as high RM turnover rates and AI workflows (now at 82% adoption) are eliminating the “Why” behind client decisions.

The Risk: When private banks like DBS or EFG hire hundreds of RMs annually, they are not only growing, they are growing Concentration Risk. If the client’s history is only stored in the RM’s mind, then the bank is nothing more than a “utility” with a limited shelf life.

The Architect’s Question: If your best RM leaves tomorrow, what does your bank actually know about your most complex clients that is not simply a scanned KYC document?

The Relationship Reality

The reality of private banking is that it is a relationship-driven business. The most advanced tech stacks in the world cannot overcome the reality that clients pay for judgment, context, and continuity. [1]

The problem private banking leaders face is that the 2024-26 business environment is highly adverse to relationship continuity. High RM turnover, aggressive acquisition strategies, and fast platform changes are creating a “Great Reset” of institutional context. We are growing our tech stacks and our teams at record rates, but we are not getting better at remembering why we made decisions in the past.

The Market Signal: Scaling Workflow vs. Scaling Judgment

The business reality is reflected in every 2026 board package:

• Hyper-Growth: Global HNWI wealth grew 4.2% in 2024. Every wealth business in the world is being asked to grow at record rates. [2]

• The Talent Carousel: DBS hired 80 RMs in North Asia in 2024. EFG hired 200 RMs globally. [3][4] The problem is that when these RMs move to competitors, they don’t just take clients; they leave behind a “Context Vacuum.”

• AI Without Continuity: While AI adoption in KYC/AML has gone from 42% to 82%, there is still a significant amount of manual intervention (6). We have improved our workflow, but we have not cured the “Dementia.”

The Case: XS Private Bank

XS Private Bank is a composite of recurring patterns found in several Tier-1 wealth franchises.

The Profile: Segment: Founder wealth and family office principals (USD $5M-$50M+ AUM).

Complexity: Multi-entity structures (Holdcos, Trusts, VCCs) and single-stock concentrations.

Expectation: “Don’t make me explain myself twice.”

The Trigger: Maya’s Discovery: The trigger event wasn’t raised by a regulator, but by a client who never raised his voice throughout the process.

The trigger event was a request to rebalance a liquidity buffer, which had been done many times before and had always been accepted by the bank.

It’s worth noting that this time around, the system treated this event like a “Day Zero” event and raised a suitability mismatch alert. It also locked the liquidity, treating a ten-year relationship like a new, unknown entity.

Maya, the Chief Operating Officer, saw the escalating issues landing on her desk. The excuses coming from the front office were always the same:

• “The old Relationship Manager approved this as an exception in 2023.”

• “We’ve always done this this way; it’s in the file.”

• “Why do we need to prove this again to the bank, which already knows?”

Maya realized the systems had performed exactly as programmed, but they had no Decision Rationale. The data migrated, but the reasoning didn’t.

The Forensic Reality: Decision Drift

The reality of forensic work is that decisions can shift over time. What broke wasn’t the policy itself, but the way the institution learns from its experiences.

XS had lost track of how to transfer three critical assets.

1. Investment Intent: Understanding the reason behind specific mandates and constraints.

2. Exception Rationale: Explaining why there was a previous risk tolerance in place.

3. The Contextual Bridge: What the bank learned last time this pattern showed up.

Whenever there’s a change in RM or the platform is upgraded, the bank’s Institutional Learning starts over from scratch.

The client remains the same, but the memory fades away.

The problem here isn’t about the RM or the platform. It’s about learning within institutions.

Why does this count as a Franchise Risk?

In private banking, trust isn’t built on rates; it comes from Continuity. When a bank forgets something important, the client stops trusting it. The client only trusts the RM.

This isn’t really a business model at all; it’s more like a Concentration Risk.

When a bank hides behind “Policy” because it’s lost its own “Judgment,” it tells the HNWI that the bank isn’t really a partner anymore, just a plain utility.

The Board’s question is simple: where exactly is memory stored?

Board responses usually mean making approvals more uniform or just buying a “better tool. “These are movements with no lasting hold.

The Board needs to ask, “If our RM walks out tomorrow and the platform gets upgraded next week, what exactly will our bank still know about our top 50 clients beyond just their KYC files?”

If the answer is just “in the files” or “in the RM’s head,” then you’re not really growing a private bank. You’re building a delicate network of personal relationships, all while managing a banking license.

A private bank dealing with Institutional Dementia just doesn’t have a clear strategy. Forgetting things regularly is one way an institution hides its weakness under the cover of being careful.

The Architect’s View

As I explained in The Foundation, the answer isn’t just better AI, it’s better Decision Engineering. We need to shift from just telling “RM stories” to building a Decision Integrity Chain (DIC) that keeps the reasons behind an exception as a lasting part of the organization, rather than just a short-term note.

Is your board dealing with what some call “Institutional Dementia,” or are you actually getting better at judging relationships as you grow? I’d be interested to know how your institution deals with the “Memory Gap” when RM transitions happen. Feel free to share your thoughts in the comments below.

Disclaimer: The cases featured in the “Boardroom Realities” series are combined stories inspired by common issues seen in various institutions and legal areas. XS Private Bank is just an example used for teaching and diagnostic reasons; it doesn’t represent any real organization. This analysis is just for information and shouldn’t be taken as legal, financial, or regulatory advice.

Footnotes

[1] Global Private Banker, Private banking is and will remain a relationship business (Karthik Shenoy, Bank of Singapore).

[2] Capgemini, World Wealth Report 2025: global HNWI wealth +4.2% in 2024; APAC HNWI wealth +4.8% in 2024.

[3] Financial Times, Singapore’s DBS courts rich Chinese with wealth manager hiring spree (80 RMs hired in 2024; at least 30 planned in 2025).

[4] Financial News London, EFGs UK boss on strategic hiring after Credit Suisse collapse (200 new relationship managers globally in 2023-24).

[5] Fenergo press release, Share of Banks Losing Clients to Poor KYC Practices Surges to Record High (client loss linked to slow/inefficient onboarding).

[6] Fenergo press release, Global Financial Institutions Struggle with Rising Client Losses and Compliance Costs as AI Adoption Increases and WealthBriefing Asia coverage (AI tools in KYC/AML 42% in 2024 to 82% in 2025; manual work remains).

[7] Dakota, JPMorgan Chase elevated David Frame to the position of global CEO of its private bank business (The company’s private banking division manages more than $2.9T in client assets globally).

During a recent discussion on core banking modernization, participants observed a trend. The following situation is an amalgamation of actual experiences of many individuals involved in the core banking transformation programs.

Two months after a major Core Banking go-live, a Tier-1 Bank found itself in an awkward position.

A long-standing mid-market client, profitable, stable, with a low default history, was automatically declined a credit line extension.

Nobody touched the file. It was blocked by the system.

Note that the new platform was cloud native, with the following key features:

• Real-time.

• API-driven.

• AI-enabled.

Everything worked exactly as designed. And that was the problem.

The Rule No One Remembered

The investigation by the bank revealed that there is a control that is hidden in the credit workflow: “Decline Automated Extensions when utilization is over 82% for more than 45 days.”

However,

• No one knew exactly how to explain 82%.

• No one could trace why for 45 days.

The rule was originally introduced in a liquidity stress episode over 15 years ago. It was never formally retired or revalidated strategically. It was simply migrated.

The case reflects a scenario where the bank improved its infrastructure but maintained its institutional memory.

And that’s the ghost in the ledger I am talking about.

Why Modernization Fails When Logic Is Not Rewritten

Typically, the core banking programs focus on

• Data inventory and mapping

• Cleansing and reconciliation

• Parallel run validation

• Phased or Coexistence Strategies

• Cloud-native

• AI-based validation tools

All necessary. All well-documented.

Your own transformation briefs may have already covered the process with these mechanics. Industry frameworks will also have covered this.

These programs typically modernize:

• Accounts

• Products

• Workflows

• APIs

• Data structures

However, they are unlikely to ponder:

• What was the rationale behind the introduction of the threshold?

• What risk event triggered this control?

• Does it still reflect the current strategy?

• Is it aligned with today’s capital regime?

• Does it support our stated customer promise?

It brings us to a clear distinction: While data migration is the act of moving data, decision migration changes assumptions.

And assumptions, taken for granted, become scalable constraints.

The Real Risk: Scalable Legacy

Across markets, banks are modernizing:

• JPMorgan is moving its retail core to Vault

• PNC is switching to FIS Modern Banking Platform

• Goldman is building Marcus as a cloud-native stack

• Temenos is integrating AI assistants using Azure

Infrastructures are changing.

However, a competitive edge will not result from a move to the cloud. It will come from how they use modernization as an opportunity to shape new decision architectures.

Programmable cores don’t just run faster; they multiply whatever rules you put into them.

As obsolete as the law is, so is the spread of the error.

Banks that reassess their decision process as they modernize experience fewer overrides, make better approval decisions, and more efficient use of capital.

Clear decisions manifest themselves on the balance sheet.

The Structural Lens: The Decision Integrity Chain™

Every central system carries a decision sequence (refer to my post on Decision Integrity Chain for details). Decisions are not made on the rule level. They begin upstream:

Purpose → Strategy → Intent → Rules → Judgment → Decision → Outcome → Feedback

Over time:

• Strategy shifts

• Risk appetite changes.

• Regulatory regimes change

• Customer expectations reset.

Nevertheless, the existing rule logic is at times still tied to past intent.

In the SME case:

• The strategy was shifting towards growth in mid-market lending.

• The liquidity profile was stronger.

• Capital buffers were strengthened.

Taking a broader portfolio level example, consider a mid-market portfolio with 12,000 SME clients.

Assuming 15% regularly operate above 80% utilization, that is 1,800 accounts governed by embedded constraint logic.

If a third of SME clients automatically trigger decline conditions at renewal windows quarterly, then approximately 600 credit decisions per quarter are being influenced by a threshold that no one can now strategically justify.

At scale, a forgotten parameter becomes a de facto capital allocation policy. The 82% rule reflected a defensive liquidity posture from another era.

The system kept the old rules. But the bank had already changed its strategy.

That gap is where integrity breaks.

Core banking modernization, when it does not realign the decision chain, carries yesterday’s caution into tomorrow’s architecture.

What a Decision Audit Actually Means

A decision audit is not data profiling. It is institutional introspection.

Before system selection and migration planning, banks should:

1. Map Decision Clusters

Not just data assets, but:

• Risk thresholds

• Override authorities

• Funding triggers

• Pricing Floors

• Exception Densities

This results in decision lineage rather than data lineage.

2. Revalidate Intent

For each critical rule:

• What was the original purpose?

• Does it align with the current strategy?

• Does it reflect today’s capital and liquidity position?

• Does it support our stated ambition regarding customer experience?

And if there is difficulty in articulating intent, the rule has moved beyond its governing context.

3. Stress Test Overrides

High override frequencies usually indicate structural misalignment. The modernization is a chance to rethink thresholds rather than to automate friction.

4. Embed feedback loops

Modern cores are programmable. Embed metrics that close the chain:

• Override Concentration

• Capital impact

• Funding sensitivity

• Customer churn due to automated deductions

This now changes the core from a rules engine to a decision learning system.

Why This Moment Matters

As AI gets ingrained within core processes:

• Automated credit recalibration

• Dynamic pricing

• Real-time liquidity sweeps

• Autonomous risk scoring

Execution authority increases.

The system will not fail loudly. It will work, but just not in the direction you think.

At scale, that is not noise. It is slow strategic damage.

The COO Question

Core banking modernization may be described as:

• A technology upgrade

• A cost optimization exercise

• A digital enablement program

It’s, in fact, a “once-in-a-decade” opportunity to align institutional logic. So, the question is not really: “Are we migrating successfully?”

It is: What layer of the Decision Integrity Chain are we modernizing?

• Only at Rules?

• Or from Purpose through Outcome?

If only the lower half of the chain is updated, the internal fragmentation inside the slower infrastructure is preserved.

And faster fragmentation is more dangerous than slow fragmentation.

The Bottom Line

The ghost in the ledger is not bad data. It’s an inherited judgment.

A core banking transformation that does not change its decision architecture does not remove legacy. It formalizes it.

The institutions that will win in this decade will not be the ones that migrate the fastest.

They will be those who audit deeply and realign their decision chain before they encode it again.

© 2026. All rights reserved. The Decision Integrity Chain™ and associated frameworks are trademarks of the author.

The Boardroom Reality: Control is Being Simulated

In 2026, most Boards believe that they are governing effectively through policy, risk appetite statements, and matrices of authority delegates.

In my previous paper, “Banking Rewired,” I wrote about the rise of a new kind of “programmable institution” in which decisions are executed through software applications rather than human judgment. This includes autonomous portfolio rebalancing, rule-based liquidity sweeps, AI-based onboarding, and real-time fraud interventions.

Execution in these environments is embedded in code. And code does not wait for committee interpretation.

When granting optimization authority to an AI agent, the Board is not delegating a task. It is delegating institutional consequences.

The machine will run precisely what’s encoded in it. Not what was assumed, what was culturally implied, and what had been discussed in the off-site.

Governance Risk of 2026 is rogue AI. It is the gap between Board intent and machine execution.

The Structural Problem: Rules Are Not Governance

Most automation programs start at the Rules level. For example.

• Credit Thresholds

• Cash sweep triggers

• Fraud scores

• Throughput Targets

While the rules are mathematical, the ‘governance’ is institutional

If institutions begin with rules and ascend to rules again, Purpose becomes merely ornamental.

And when Purpose becomes merely decorative, then Boards are in charge of optics, not outcomes.

This is what I define in my recent SSRN paper, AI Optimizes. The Bank Pays the Price, as the Fiduciary Gap, the structural divergence between what an AI system is optimized to do and what the institution assumes it is protecting.

Optimization is measurable. Stewardship must be encoded into execution.

The Control Architecture: The 8 Layer Decision Integrity Chain™

The Decision Integrity Chain (DIC) is not an ethics framework.

It is a control architecture that links Board purpose to autonomous execution, which is auditable, reversible, and capital-aware.

Every autonomous action must pass through eight linked layers.

1. Purpose - Why We Exist

That is the rationale behind the institution.

Before we automate anything, we should ask: Does this action fit who we are and what we stand for?

If we cannot defend it at that level, then it should not be automated. From there, the process of holding people accountable begins

2. Strategy - How We Win

This is where our competition happens and where our money goes.

Rather, automation needs to help us build on our position instead of quietly undermining it by focusing keenly on specific objectives.

If the machine improves one number but decays our original position, then we are not being helped.

3. Intent - What We Are Protecting

For each type of decision, we have to be clear about: What exactly are we trying to protect?

Not just profit or speed, but long-term stability and trust.

This is where the actual risk comes in, namely, the disconnect between what the machine is optimizing and what the institution should be optimizing.

4. Rules - If Then Logic

These are the triggers and thresholds that are set by the system.

Rules are necessary. But rules are not wisdom.

Furthermore, the system will probably reach technically sound but institutionally harmful conclusions as long as it relies on the decision rules.

5. Judgment - Veto Layer

This is the safeguard - Where is qualitative reasoning encoded?

If the potential financial gain to be realized does not measure up to the cost of fiduciary erosion, then the system fails to execute.

Judgment is where stewardship becomes executable.

6. Decision - Commitment Layer

It’s the moment of authorization. To be a legitimate autonomous decision, it must pass the judgment layer.

Without judgment, the decision is automation. With judgment, decision is governance.

7. Outcome - Institutional Layer

Now we look at the reality.

• What actually happened at the Franchise level?

• Did we gain some efficiency but lose some trust?

• Did capital seem safer but become less mobile?

• Did fraud decrease when customer friction increased?

Outcomes are more important than institutional metrics.

8. Feedback - Governance Loop

This is bridging the gap between Purpose and Outcome.

Thus, when the regulator or the large shareholders say, “Why did this happen?”, it should be traceable sequentially at all eight layers.

If it cannot be traced, then alignment is assumed, not demonstrated.

Where the Chain Is Breaking in 2026

Every major boardroom debate today reflects a fracture somewhere in the Decision Integrity Chain.

These are not isolated operational issues. They are discontinuities in institutional decision flow. For example, in the Banking:

When the chain fractures, optimization accelerates misalignment.

The Board Diagnostic

Before making any significant autonomous decision class approval or renewal, a Board should require answers in writing to the following questions:

1. How does this system reinforce our institutional Purpose?

2. Does this automation improve or modify our strategic and capital posture?

3. Is the fiduciary objective explicitly encoded beyond yield or speed?

4. Are the rules transparent, owned, and explainable

5. Where is veto power held, and is it pre-written or made-up?

6. Is execution reversible, and with what level of materiality?

7. Are we measuring the impact of the franchise itself, rather than just the

8. Can any material outcome be sequentially traced back to Board intent?

Unless these questions can be answered readily and in writing, optimism has displaced governance.

The 2026 Mandate

AI will execute. But the question is, does execution reflect institutional intent, or does it reflect only computational efficiency?

The Decision Integrity Chain is not an IT framework. It is a fiduciary instrument.

Governance in an agentic institution is no longer the policy oversight of. It is an architectural design.

And architecture does not drift silently.

© 2026. All rights reserved. The Decision Integrity Chain™ and associated frameworks are trademarks of the author.

In January 2026, the world’s Global Systemically Important Banks (GSIBs) report a historically “bulletproof” level of capital. JPMorgan Chase again leads the pack with a fortress balance sheet, while HSBC has retained its CET1 ratio of around 14.6% following its structural reorganization in 2025.

However, in the boardrooms, a new air of unprecedented constraint prevails. The reality of 2026 is that a bank with a record amount of capital is “fragile,” and that is because its capital is no longer “maneuverable.”

The Localization Trap

For fifteen years, the “Basel III” dream was a harmonized global playing field. In 2026, that dream ended and was replaced by Geopolitically driven Fragmented Globalization.

• The US vs. EU Divergence: After a regulatory shift in late 2025, US regulators are now pursuing a more innovative, “capital-neutral” Basel III Endgame. Meanwhile, the EU and UK are doubling down on Basel 3.1, with the adoption of “hard” output floors that constrain the ability of internal models to reduce capital requirements.

• Real-World Signal: A corporate loan in Paris will now require substantially more capital than the same loan in New York, requiring banks to cope with “islands of capital” rather than a global pool.

• Board Question: “Are we running a global bank, or a group of locally capitalized islands?”

Metrics for a Fragile Reality

Boards are no longer content with a single CET1 ratio. In 2026, they monitor three other forces:

(a) Bucket Sensitivity

In the 2025 GSIB review, Bank of America increased a systemic bucket, imposing a multi-billion-dollar capital buffer. Deutsche Bank cut its exposures to move down.

This is not window dressing. A bucket move changes strategic capacity itself. Growth is now accompanied by a capital step-function risk.

Boards now ask: Are we maximizing the franchise or operating around a threshold?

(b) The Shadow Backstop

Institutions like Goldman Sachs and BNP Paribas are significant liquidity suppliers to the burgeoning private credit market.

On the balance sheet, banks display diversified exposure. In times of stress, correlations rise.

If parts of the private credit market fail, banks are not bystanders. They are counterparties themselves. Excess capital on the balance sheet may actually support levered non-bank business models.

Board Question: Is our buffer really defensive, or is it effectively used as first-loss capital for shadow banking?

(c) Trapped Equity and ROE Drag

Capital trapped in low-growth or heavily regulated countries imposes ROE drag. Treasury departments look for balance sheet workarounds. But pain is evident:

• Group ROE suppression

• Heightened transfer pricing disputes

• Capital suboptimality masked by earnings smoothing

The problem is no longer sufficiency. It is usability.

Simulating Fragility

Supervisors are now requiring institutions to go beyond historical stress testing.

The ECB’s 2026 guidance requires supervised banks to assess geopolitical risks with the potential to cause at least a 300-basis-point CET1 depletion. This is not incremental volatility; this is capital shock velocity.

Boards are now using balance sheet “digital twins” to assess the following risks:

• Trade embargo spillovers

• Sudden sanctions regimes

• Equity market drawdowns

• Funding freezes

The focus is no longer: “Do we meet the minimum capital?”

The focus is now: “What buffer will be depleted first?”

The AI Capital Overlay

Under emerging US regulatory guidance related to the GENIUS Act, autonomous AI systems could be subject to further operational risk capital charges.

The implication is subtle but significant:

• Automation reduces costs.

• Autonomy increases model risk concentration.

• Model risk concentration increases capital intensity.

The productivity benefit of AI is partly offset by the complexity premium.

Capital planning now includes technology architecture.

The 2026 Reality Check

A 15% CET1 ratio does not imply the bank is resilient. In 2026, fragility will be characterized by:

• Capital being geographically trapped

• Counterparty risk with layered non-bank leverage

• Strategic growth with bucket-constrained potential

• Operational complexity absorbing incremental buffers

What It Looks Like vs What It Feels Like

To The Business

• Money moves more slowly

• More time spent fixing internal funding issues than driving growth

To The Board

• More stress test discussions

• More calls with regulators

• More discussions of what if this freezes

To Revenue

• Returns decrease

• Group ROE decreases due to stuck capital that cannot be used efficiently.

The Boardroom Question

“If our capital is legally trapped in one region and we are providing liquidity support for shadow banks in another region, do we really have a capital buffer, or are we just maintaining an extremely costly and immobile vault?”

The Accountability Gap Under AI Scale

In the first two notes of this series, we looked at a loan approval and at deposit stability. In both cases, the systems behaved exactly as designed

• The loan is extended just as the system is intended to extend it.

• The deposits behave in a perfectly logical way for the customers.

The trouble starts when the board asks a simple question, and the organization hesitates.

The credit question: Why did we think this customer would remain safe?

The liquidity question: Why did we think these balances were stable?

This note examines the pattern beneath both. It cuts across credit, deposits, compliance, and fraud.

The pattern emerges when decisions begin to outrun the organization’s capacity to understand them: a governance gap.

This creates structural asymmetry inside the decision architecture.

The Decision Integrity Chain © under AI

In the traditional model, there was a clear chain of events:

• Purpose was used to define risk appetite.

• Strategy was used to define operating boundaries.

• Intent was used to clarify purpose.

• Quantitative rules were used to set measurable limits.

• Qualitative rules were used to interpret context.

• Judgment was used to interpret signals.

• A decision was reached.

• An outcome was generated.

• Feedback was used to inform future decisions.

The strength of this system was not its speed. The strength was its coherence along this chain.

In this new world of AI scale, one node has accelerated significantly: The Decision node is now automated, distributed, and instantaneous.

However, three other nodes have quietly degraded:

• Judgment is partially replaced.

• Qualitative rule interpretation is condensed.

• Feedback is often delayed or incomplete.

Most importantly, there has been no redesign of the Reconstruction layer along this entire chain.

The chain is still intact. However, it is no longer symmetrical.

When the Reconstruction layer is weakened:

• Accountability is blurred.

• Legal defensibility is compromised.

• Board oversight is reactive.

• Trust is undermined silently.

The institution believes this chain is intact. But it is not.

It is operating at decision speed with governance latency.

The signal: When speed becomes the risk

Across major banking markets, the debate on whether to use AI is over. Today, it is about scale.

• Straight-through credit is now the norm for many retail underwriting operations.

• Real-time fraud decisions operate in millisecond timescales to match the instant payment rails.

• Customer exit decisions are increasingly being driven by automation through AML, sanctions, and behavioral models.

These are no longer pilots. They are production systems integrated into core operations.

From an operations perspective, it looks like they are successful. They are cheaper. They are near instant. And they are reducing exceptions.

But behind the scenes, the bank is accumulating something else, called explanation debt.

The gap: Decision speed vs Governance speed

Banks are scaling AI in high-stakes decisions. While Governance frameworks are being scaled but not the Judgment reconstruction. This creates structural asymmetry inside the decision architecture.

1. The ROI Paradox

Efficiency is easy to measure. Audit trails are slower.

When a decision is made through the interaction of multiple models, understanding what actually happened is slow.

While the decision was fast, the explanation is slow.

2. The Black-Box Exit

When a loyal customer is terminated by the automated system, the bank can explain the control that initiated the action. However, the bank struggles to clearly explain why it happened to that customer, at that time, in that context.

That is the point at which legal and reputational risks are born.

3. The Accountability Blur

When an automated fraud detection goes wrong, the initial response is: “The system decided.”

No one owns it. It is someplace between data, risk, technology, vendors, and policy. The controls may have worked. No one owns the judgment.

The speed of the system makes it difficult to point to who is accountable.

The Reconstruction Test

Imagine the following scenario.

A customer who has banked with the institution for 18 years is suddenly told that the bank is terminating the relationship.

• No previous complaints.

• No missed payments.

• No public issues.

The termination was caused by a combination of behavioral flags, sanctions screening correlations, and a network risk model update that went live two weeks ago.

The media calls. The board asks to reconstruct.

Three teams review the logs.

• Risk team: We can see the control was fired correctly.

• Data team: We can see the model version was validated.

• Operations team: We can see policy thresholds were met.

All of this is technically correct.

But when asked the simple question: Why this customer, on this day, under these circumstances?

The teams can show that a decision was made. But no one can explain the rationale for the decision.

That is the governance gap.

The strategic divide: Speed vs Clarity

There is a split in the way banks are approaching.

• Some are optimizing for speed. They are focused on speed in their decisions, their coverage, and their cost.

• Others are optimizing for clarity. They are focused on explaining what happened and why.

This is leading to a structural split.

One bank may demonstrate that a control worked as expected. Another bank may walk someone through a decision step-by-step and explain their reasoning at that time.

They may both be in compliance. Only one of them will be defensible.

Compliance is about: Was the control present?

Defensibility is about: Can we clearly explain why we made that decision?

What leading institutions are building

Leading banks are building a missing layer: Decision trace architecture ©

The architecture is not just about model logs or policy documentation. It incorporates actual “decision trace payloads” embedded inside high-risk business workflows:

• Decision owners assigned to specific model outputs

• Time-stamped rationale layers at the point of override

• Model interaction maps, highlighting contributing signals

• Version-level traceability to specific customer outcomes

• Accountability assignments at the decision node level

They’re building this so that every high-stakes decision can be reconstructed without assembling three committees after the fact.

A metric every Board should track

Boards may need to track a new metric: Decision Reconstruction Time (DRT) ©

How long does it take to fully reconstruct a high-stakes AI-assisted decision in a way that is legally defensible and regulator-ready?

• If the answer is weeks, the institution has an explanation debt.

• If the answer is hours, governance design is embedded.

Speed of execution without speed of reconstruction is a latent risk.

Supervisory signal

Recent supervisory communications across major jurisdictions increasingly focus on model governance, accountability mapping, and explainability in customer-impacting AI systems.

The shift is subtle but important.

The question is no longer only: “Was the model validated?”

It is increasingly: “Can senior management clearly evidence why this decision was appropriate at the time it was made?”

That is a higher bar.

The question before the Board

If an AI-assisted decision goes wrong, and no human was present at the time, who is liable?

It is not the model or the policy. It is the person.

Until that person can be identified, quicker decisions will continue to flow to the board long after the logic behind them can be deciphered.

Speed without explanation is not innovation. It’s delayed accountability. And this isn’t a tech problem. It’s a governance design problem.

That’s the missing layer. And that’s now a risk to the board.

The shift beneath the surface

Over the past two decades, the Deposit franchise of a bank has transformed across all six key dimensions, namely:

• Product design,

• Dynamic product pricing,

• Providers (new providers entered through BaaS and fintech partnerships),

• Channels (mobile and platform-based distribution),

• Customers (more informed, more connected, and more rate sensitive), and

• User experience.

The transformation has offered significant rewards.

However, the rewards have been delivered with a very high degree of risk, both for customers and franchisees. These risks are spread across all four dimensions of liquidity, operations, reputation, and systemic.

The deposit runs of 2023 exposed four key risks:

• The speed at which customers can move funds,

• Rate of increase in the concentration of uninsured balances,

• Behavior of digital-only and rate-sensitive deposits under stress,

• Opacity of depositor profiles originated through platforms or marketplaces.

The runs presented structural weaknesses at banks in balancing risk with customer demands. With regulators emphasizing higher liquidity and customers actively seeking better returns, banks are facing significant challenges to grow their Deposit franchise.

This note illustrates how that tension is playing out inside a bank today for its Deposit franchise.

The case is a composite scenario, based on observable patterns across large Asia-Pacific banks with significant Singapore operations. It does not refer to any specific institution.

Consider a leading regional bank delivering strong business performance, including stable deposit balances. Following the 2023 deposit runs, it strengthened liquidity buffers and governance oversight. On paper, the franchise appeared resilient.

In August 2024, during a routine review of third-quarter funding and liquidity metrics, a senior executive made a comment that sounded entirely reasonable at the time:

“We have been able to manage our Deposits business. Pricing is the only real issue.”

The Deposits portfolio looked good.

The top 20 Depositors accounted for approximately 18% of the total.

Liquidity Ratios were healthy

• Liquidity Coverage Ratio (LCR) was approx. 135-140%, well above the regulatory requirement.

• Net Stable Funding Ratio (NSFR) was 115–118%, above the required thresholds.

Stress scenarios showed resilience under standard outflow assumptions.

There were no red flags. On paper, this looked like a balanced regional deposit franchise behaving exactly as expected.

A quick look at the bank’s funding model

The model was based on three clear assumptions.

Q3-2024 - the signal that drifted

Shortly after the August 2024 review, the bank began experiencing a shift that didn’t reflect in the dashboards, though. Interestingly, the shift did not start in retail, unlike the deposits run. The change began in SME and upper-mid corporate operating accounts.

Between September 2024 and February 2025, the parameters’ performance was:

Total deposits were stable. The mix had changed, though. Nothing alarming.

But clients were moving the money

This time, it was not a deposit run. It was a fund optimization strategy by the firms.

The client firms were not distressed. They were being rational.

Policy rates were rising, and money market funds were passing through higher yields more quickly than traditional operating accounts. In some cases, excess corporate cash could earn 100–150 basis points more outside standard bank deposits.

At the same time, treasury platforms and bank APIs made same-day sweeps into money market funds operationally seamless. What had once required manual treasury action was now automated. And idle balances were being seen as inefficiency.

Soon, several clients moved idle money away from their bank accounts. This was prudent liquidity management from the client’s perspective

The bank did not lose clients. The accounts were active, and transactions were flowing. But it invalidated the bank’s core assumption that operating activity implied funding stability.

Despite balance migration, the issue did not trigger formal escalation

Over the 5 months, the following represents the movement of key parameters performance.

As can be observed, no regulatory thresholds were breached. LCR and NSFR were within limits, and internal liquidity dashboards reflected the shift in mix.

Based on dashboards, the Treasury categorized the movement as competitive pricing pressure. But Business teams saw active client relationships while Risk systems didn’t show any deterioration.

However, there were four consequential shifts, which were not reflected in dashboards:

• Even though aggregate balances remained stable, effective duration assumptions declined by roughly 40–50%. Modeled behavioral stability, which is derived from rolling balance decay patterns, was compressed over two quarters.

• At the same time, operating balances were funding a shrinking proportion of their associated lending exposure, falling below 60% in certain SME segments. The residual gap was being bridged through higher-cost wholesale funding.

• The economic substitution cost was visible in internal transfer pricing, but it was treated as competitive repricing pressure rather than a structural funding shift.

• Importantly, regulatory classifications did not change. SME operating balances continued to receive stable funding treatment under prevailing liquidity assumptions, and no supervisory thresholds were approached.

At the time, the prevailing view of the team was that these movements reflected cyclical rate dynamics rather than a durable change in depositor behavior, while the signals were present across treasury analytics, FTP reports, and balance composition trends. Clearly, while data was not the problem, its interpretation was.

In every monthly and quarterly review, after the discussions, the conclusion was reassuring: “This is pricing pressure, not a structural issue.”

And it stopped there.

The discrepancies that surfaced in April 2025

However, by March–April 2025, the reality was surfacing.

For the April’25 Board meeting, the packs showed

• The funding costs have increased faster than internal forecasts, despite stable deposit balances.

• Net interest margin compression has exceeded, which could not be explained in view of the rate movement.

• Liquidity stress testing did not provide a comfortable explanation.

• Supervisors were asking more pointed questions about the stability of operational and uninsured deposits, particularly in SME and corporate segments.

• It was interesting because, individually, none of these developments breached approved limits. Collectively, they created tension between reported stability and economic performance.

Therefore, the board requested a structured funding review.

A subsequent deep dive by the team confirmed that

• There is an underlying shift in the deposit mix, while Headline deposit totals remained intact.

• Retail balances were stable.

• However, SME and upper-mid corporate operating balances had declined in effective stability and were increasingly being replaced by higher-cost wholesale funding.

The issue was no longer deposit loss. It was funding quality.

The question was no longer “are we losing deposits?”
It was “Are we still funding ourselves the way we think we are?”

The team, based on the findings, prepared a proposal for the Deposits transformation. The proposal included various aspects aligned with market trends like the emergence of Platform banking with deposits integrated into non-banks, Cash sweep networks, deposit marketplaces, and the growth of BaaS (Banking-as-a-Service).

Board mandate in July 2025 and outcomes

By May 2025, the funding review moved from management discussion into ALCO (Asset-Liability Committee) and ultimately to the board.

The debate had shifted. The question was no longer whether deposits were stable in aggregate. It was whether the bank understood the behavioral foundations of that stability.

During the meeting in July 2025, the board asked a simple but uncomfortable question: Are we still defining “core deposits” correctly?

The board did not approve the transformation program proposal. Instead, it asked the team to implement three contained, time-bound mandates designed to test specific funding assumptions and inform strategic choices. These being:

• Behavioral Segmentation of Deposits

• Liquidity Stress Recalibration

• Early Warning Enhancements for funding shift

What it meant for the team and the outcomes:

A joint team across Treasury, Product, and Risk formed a task force to examine the deposit franchise. The task force. amongst others, focused on

• Retention patterns across 30-, 90-, and 180-day horizons

• Rate sensitivity

• Sweep activity into money market funds and external platforms

• Uninsured versus operational classifications

• Platform-originated versus relationship-originated flows

• Stress scenarios against existing liquidity buffers and wholesale funding capacity.

By December 2025, the results were clear.

• In selected SME segments, up to one-third of balances behaved more like tactical liquidity management than stable operating funding.

• No regulatory thresholds were breached. The bank remained compliant

• Funding resilience proved more sensitive to behavioral shifts than earlier models.

• Behavioral shifts were being detected earlier in the funding cycle, particularly in SME segments exhibiting rising rate sensitivity.

The insight, however, was strategic rather than technical.

The bank had the option to defend rate-sensitive balances through higher pricing and margin compression. But it would increase reliance on wholesale funding.

The trade-off seemed structural.

The matter returned to the board in January 2026

Despite the board’s mandate and the subsequent analytical work, funding economics did not normalize.

• Incremental wholesale funding costs rose from approximately 3.6% to above 4.4% within two quarters.

• Asset repricing helped to partly offset the costs

• Margin compression continued.

By January 2026, at the first full board meeting of the year, the funding outlook was revisited under the revised stress framework. The updated scenarios incorporated higher behavioral velocity, elevated deposit betas, and greater migration risk through sweep channels.

The discussion shifted.

The issue was no longer analytical sufficiency. The data was complete, and the teams understood the data. The models were also recalibrated, and monitoring had improved.

The question was interpretive and strategic.

If deposit behavior had changed structurally, maintaining prior balance levels would require paying materially more for those balances.

That, in turn, would alter relationship economics, lending appetite, and internal transfer pricing assumptions.

This was not a treasury calibration. It was a deposit franchise choice.

One board member articulated the tension directly:

If this is structural, which balances are we prepared to pay up for, and which are we willing to lose?

That question reframed the discussion. The debate moved from dashboard metrics to capital allocation and competitive positioning.

It was no longer about whether funding was compliant. It was about what kind of funding model the bank wanted to run.

Why this is now a boardroom reality

Until now, the bank had treated deposit behavior as:

• a pricing issue,

• a treasury issue,

• or a modelling issue.

In January 2026, it became something else. It became a governance issue.

The board was no longer asking:

• Are deposits leaving?

• Are ratios compliant?

They were asking:

• Do we still understand the economic logic of our funding model?

• And if customer behavior has changed permanently, which assumptions no longer hold?

That is the reality check. While nothing had collapsed, the Headline Deposits remained almost the same, but the assumptions that defined “core funding” were no longer reliable.

The bank no longer had a visibility problem. It had a decision problem.

• Matching market pricing would compress margins.

• Allowing balances to run off would increase reliance on wholesale funding.

• Tightening liquidity buffers would reduce lending flexibility.

For the Board, it was not a crisis. It was about a structural choice about the future of the deposit franchise.

The question before the board

By this stage, the debate had narrowed to a single issue.

If deposit behavior had changed, which funding assumption no longer held, and who was responsible for revisiting it?

This was not a treasury calibration. It was not a pricing adjustment. It was a governance decision.

If behavioral velocity proved structural, then “core deposits” could no longer be defined purely by product classification. Stability would have to be assessed based on observed behavior.

The question before the board was no longer whether deposits were leaving.

It was whether the bank was prepared to redesign its funding model around a different definition of stability.